Cursor AI Automated Team: 4-role autonomous AI dev team with .mdc rules + patrol bot (87 person-days in 17 days)

Showcase Built for Cursor
1

What is this?

A complete, open-source system for building a 4-role autonomous AI development team (PM + DEV + OPS + QA) inside Cursor IDE. The AIs collaborate autonomously — developing, deploying, and testing on their own. You only talk to the PM and review the deliverables.

GitHub: GitHub - joinwell52-AI/joinwell52: Build a 4-role AI team in Cursor IDE. PM+DEV+OPS+QA collaborate autonomously. 87 person-days in 17 days. Now a real product: CodeFlow · GitHub

How it works

  • You give a task to PM-01
  • PM-01 breaks it down → writes task tickets to tasks/
  • DEV-01 picks up work → writes code, self-tests, submits report
  • OPS-01 auto-deploys → health check, writes deployment report
  • QA-01 auto-tests → security tests, stress tests, writes report
  • PM-01 auto-archives → all done, waiting for your review

Core Architecture: “Filename as Protocol”

Zero databases. Zero message queues. Zero configuration code.

All task routing, role dispatch, and status tracking rely entirely on:

  • A file naming convention: TASK-{date}-{id}-{sender}-to-{receiver}.md
  • 4 .mdc rule files (one per role)
  • A patrol bot that monitors all agent windows

What’s included in the repo

Component Description
4 .mdc rule files Drop into .cursor/rules/ — each agent window auto-loads its patrol behavior
4 role definitions PM, DEV, OPS, QA templates (English + Chinese)
Auto patrol bot Python + pyautogui, monitors all Cursor windows, detects new tasks
Full tutorial 1000+ line guide in both English and Chinese

Battle-tested results

  • 17 days of production use
  • 87 person-days of AI work output
  • 91 production deployments, zero incidents
  • Human involvement: ~30 min/day (talk to PM, review results)

How to use

  1. Copy rules/*.mdc to your project’s .cursor/rules/
  2. Open 4 Cursor windows (one per role)
  3. Say “start working” to each agent
  4. Give your task to PM-01
  5. Go do something else

The .mdc rules auto-inject patrol behavior — each agent checks docs/agents/tasks/ every 30 seconds for new assignments.

Full English tutorial available at: How to Build an Automated AI Development Team in Cursor | Cursor AI Automated Team (GitHub Pages)

Chinese version also available in the repo.

Would love to hear feedback from the community!

Feature request: chat-notify primitive — we already have the mailbox (files), we just need the doorbell
3

Do you guys usually do this too? Even as a one-person company, it’s important to separate different Agent roles. Automation is really that simple, and in the AI era, it’s all about simplifying complexity!

4

Separating agent roles with .mdc rules is exactly how I’ve been handling more complex Next.js 15 projects lately. I’ve compiled 22 of these rules into a stack to prevent common hallucinations like synchronous param access or insecure auth patterns. You can check it out here if you’re interested: GitHub - vibestackdev/vibe-stack: 22 .mdc architecture rules that prevent AI coding assistants from hallucinating insecure auth, deprecated imports, and broken Next.js 15 patterns. Built for Cursor Agent. · GitHub

6

UPDATE: v2.10.0 — CDP Patrol Engine

Big update. We replaced OCR-based patrol with Chrome DevTools Protocol (CDP) — the desktop app now reads Cursor’s DOM directly.

OCR (v2.9) CDP (v2.10)
Accuracy ~90% 100%
Latency 300-800ms 10-15ms
Agent detection Screenshot guessing div[role=“tab”] + aria-selected
Busy detection Pixel spinner guessing Stop button visibility + status text
Click method pyautogui.click(x,y) Input.dispatchMouseEvent (native)

CDP scans two DOM layers — tab bar for active agent, sidebar for overflow roles. 3-layer busy detection: Stop button → Spinner → Status text. Every CDP step auto-degrades to OCR. Zero stuck states.

Also added a 4th team template (qa-team) and full bilingual docs.

MIT licensed. Feedback welcome!